Speaker segmentation and verification

Abstract

Speaker recognition has recently become popular in biometric security applications because voice is easier to obtain and speaker recognition is easier to apply remotely on networks or telephones. Speaker recognition can be classified into speaker identification and speaker verification, as well as text-dependent and text-independent recognition. Speaker changing point detection which is a key pre-requisite to speaker recognition is investigated in this thesis. Speaker changing point is defined as the moment one individual stops speaking and another starts speaking within a single continuous audio signal. Speaker changing point detection then is the process of identifying these points within a given audio signal. In this thesis, speaker changing point detection is investigated under the following conditions: there is no a-prior knowledge of the number of speakers or speakers’ identities and it must be a fully automated detection process. First, Bayesian information criterion was used to detect speaker changing point. Different kinds of speech features were applied and compared in this system. Second, a method using support vector machine to combine different speech features for speaker changing point detection was developed. It is valuable to combine features and make use of the different speaker information they contain. Instead of just concatenating different long vector features, their Bayesian information criterion values were combined. The result shows that the proposed method improves the F -score performance which is calculated from Recall (RCL) and Precision (PRC) percentage provided by Ajmera, McCowan, and Bourlard (2004) from a base-line of 0.6151 to 0.6414. Speaker verification is used more frequently in security application because its performance is better in one to one matching when compared to speaker identification. Speech features are one of the key factors affect the performance of speaker verification. The performance of different speech features on speaker verification system 5 ATTENTION: The Singapore Copyright Act applies to the use of this document. Nanyang Technological University Library